Recently, Chen Lei, a master student with the School of Electronics and Information Engineering(SEIE) of IMU who majors in control engineering published his 2 research results respectively at ICME 葡萄官网入口, an international top conference in the field of multimedia(Class B conference recommended by CCF) and in IEEE Transactions on Circuits and Systems for Video Technology. The two researches were all completed under the guidance of Associate Professor Dai Huhe with SEIE and Associate Professor Zheng Yuan with the College of Computer Science.
The full name of ICME is IEEE International Conference on Multimedia and Expo, one of the most important academic conferences in the field of multimedia. Chen’s article that has been accepted by ICME is titled “ICANet: A LIGHTWEIGHT INCREASING CONTEXT AIDED NETWORK FOR REAL-TIME IMAGE SEMANTIC SEGMENTATION”(Oral Presentation). In this article, the author proposes Increasing Context Aided Network(ICANet). Considering the segmentation accuracy differences existing in low-parameter network, the author proposes Inverted Depthwise Sparable convolution block(IDS block) which can extract rich semantic information from the feature map with few channels. On the basis of IDS block, ICANet encoder is constructed. In ICANet encoder, the strategy of across-resolution fusion is used to efficiently fuse the feature output of encoder. Specifically, proposed ICANet with only 0.56M parameters obtains 72.4% mIoU at 178.2 FPS and 76.1% mIoU at 221.1 FPS on Cityscapes and CamVid test sets for full-resolution inputs, respectively.
Chen Lei delivers his academic speech at ICME 葡萄官网入口.
The article that has been accepted by IEEE Transactions on Circuits and Systems for Video Technology is titled “RAFNet: Reparameterizable Across-resolution Fusion Network for Real-time Image Semantic Segmentation” (DOI: 10.1109/TCSVT.葡萄官网入口.3293166). In the article, the author proposes Reparameterizable Across-resolution Fusion Network(RAFNet) for real-time image semantic segmentation. RAFNet adopts the encode-decoder structure. The author also proposes Reparameterizable Channel & Dilation convolution block(RCD block) which uses re-parameterization to decouple training RCD block and inference RCD block to have higher segmentation accuracy and faster inference speed. Moreover, RCD block uses the dilated convolutions with different dilation rates to extract feature information from wide receptive fields. On the basis of RCD block, RAFNet is constructed. To have a higher segmentation accuracy, across-resolution fusion strategy is proposed for RAFNet decoder whose high fusion efficiency arises from the encoder output at different stages. Specifically, the proposed RAFNet with only 0.96M parameters obtains 75.3% mIoU at 107 FPS and 75.8% mIoU at 195 FPS on Cityscapes and CamVid test sets for full-resolution inputs, respectively. .
The first author of the above-mentioned articles is Chenlei. The two researches are supported by the National Natural Science Foundation of China, the Natural Science Foundation of Inner Mongolia and Inner Mongolia University Program for Launching Research of Introduced High-Level Talents
